re-convnext

A re-implementation of ConvNeXt in PyTorch as described in

A ConvNet for the 2020s. CVPR 2022.\ Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell and Saining Xie\ Facebook AI Research, UC Berkeley

What is this about?

The official repository contains the final version of the model (dubbed ConvNeXt) after all the modernizing steps described in the paper, whereas we provide a standalone model for each checkpoint (macro design, resnext, inverted bottleneck, etc.).

This repository mainly serves an introductory and educational purpose as we document the implementation progression from ResNet to ConvNeXt in PyTorch.

Explanation

Here is the modernizing roadmap from the original paper:

There are 7 checkpoints in the implementation roadmap: ResNet, Macro Design, ResNeXt, Inverted Bottleneck, Large Kernel, Micro Design (ConvNeXt).

The implementations are in models folder in the following order:

• We started with torchvision's ResNet, simplified it to resnet.py, integrate stochastic depth in resnet_stodepth.py.

• Added the macro changes described in the paper, resulting in macro_changes.py.

• Added the ResNeXt-like changes described in the paper, resulting in resnextify.py.

• Added inverted bottleneck in inverted_bottleneck.py.

• Moved depthwise conv. up and use larger kernel in large_kernel.py.

• Added micro-design in micro_design.py. This is already equivalent to ConvNeXt.

• Cleaned up the code to get the final checkpoint in convnext.py.

In each checkpoint, the changes made with respect to the previous checkpoint, except for convnext.py, are documented in the code in NOTE comments, e.g.

# NOTE: Inverted bottleneck
expand_width = planes * self.expansion
self.conv1 = conv1x1(inplanes, expand_width)
self.n1 = norm_layer(expand_width)
self.conv2 = conv3x3(expand_width, expand_width, stride, depthwise=True)
self.n2 = norm_layer(expand_width)
self.conv3 = conv1x1(expand_width, planes)
self.n3 = norm_layer(planes)

References

• The ConvNeXt paper which this implementation is based on.
• The official implementation, where we double checked to make sure the final checkpoint convnext.py is equivalent to ConvNeXt.
• Torchvision's resnet.